Skip to content
Leonid Kuznetsov

Free learning path

The Path From Frontend to AI Engineer

I did this transition while shipping production software, and most roadmaps I read at the time were either academic ML curricula or lists of tools with no order. This is the path I would give a strong frontend engineer today: eighteen phases, what to learn in each, what to build to prove it, and exactly which resources cover it.

Every link below was opened and checked. None are affiliate links, and I have no relationship with any platform listed. Where nothing good exists on a topic, the phase says so rather than sending you somewhere that will not help. Your progress is stored in your own browser and never leaves it.

First, the job

What an AI engineer actually does.

Most confusion about this transition comes from conflating three different jobs. This is the one that hires frontend engineers.

The job

  • You build products where a model is one component among many — retrieval, tools, state, permissions, fallbacks, and a UI that survives a bad response.
  • You are measured on whether the feature is reliable, affordable, and shippable, not on model accuracy in isolation.
  • You own the parts nobody else wants: the failure paths, the evaluation harness, the cost line, and the answer to 'why did it say that?'
  • You spend more time on data, context, and interfaces than on models.

Not the job

  • You are not training foundation models. That is a different job at a handful of companies.
  • You do not need a PhD, a research background, or CUDA.
  • You are not a prompt engineer. Prompting is one skill in the stack, not the stack.
  • You are not replacing your frontend experience — you are compounding it.

Your starting position

What frontend already gave you.

You are not starting from zero, and the parts you already have are the parts AI teams are usually missing.

  • You already build against unreliable, high-latency, streaming APIs — which is exactly what a model is.
  • You already think in loading, empty, partial, and error states. AI features live or die in those states.
  • You have product instincts. Most failed AI features are product failures, not model failures.
  • You ship to real users under real deadlines, which is the scarce skill on AI teams full of notebooks.
  • You know TypeScript, and the current generation of AI tooling is increasingly TypeScript-first.

Before you start — what employers actually ask for

I read 68 AI engineering job postings so this path targets real requirements.

The companion study covers application-layer AI Engineer roles across frontier labs, AI startups, and SaaS scale-ups — which languages and tools they name, what the ideal candidate looks like, and what the job pays. Research and model-training roles were excluded on purpose, because they are a different job with different skills.

Read the job-market study →

Free start

You can begin today without paying anything.

Every phase below has at least one free resource. These are the ones worth opening first — they are genuinely competitive with the paid alternatives.

Practical Deep Learning for Coders (opens in a new tab)

Jeremy Howard · fast.ai

Machine learning taught top-down for people who already program. The accompanying book is free online too.

Neural Networks: Zero to Hero (opens in a new tab)

Andrej Karpathy

Builds a language model from scratch on video. The single best way to stop treating LLMs as magic.

Practical Python (opens in a new tab)

Nina Zakharenko

A complete written Python course with exercises — the free companion to the Frontend Masters course, and the fastest way into the language if you already program.

Hugging Face Learn (opens in a new tab)

Hugging Face

Open courses on LLMs, agents, MCP, and NLP. An account is needed only for optional certificates.

LangChain Academy (opens in a new tab)

LangChain

Free courses on LangGraph and agent orchestration. Signup required, no payment.

LLM Zoomcamp (opens in a new tab)

DataTalks.Club

A full open-source course on building an LLM system over your own knowledge base, homework included.

AI Eval Flashcards (opens in a new tab)

Hamel Husain & Shreya Shankar

Condensed takeaways from the best-known evals course. Email signup, no payment.

DeepLearning.AI courses (opens in a new tab)

Andrew Ng · DeepLearning.AI

Short courses on prompting, RAG, and agents. Videos are free to watch; labs and certificates are not.

Frontend Masters workshops (opens in a new tab)

Frontend Masters

Live workshops are periodically free to attend with an RSVP, even without a subscription.

The roadmap

Eighteen phases, in two tracks.

Each phase has a goal, skills you can check off as you go, one thing to build that proves the phase, and the resources that cover it. Where nothing good exists, the phase says so. Tick the boxes as you work — your progress stays in this browser.

Your progress

0%

0 of 81 skills · 18 phases

Saved in this browser only. Nothing is sent anywhere, and there is no account to create.

Engineering foundations

0/35 skills

The base an AI engineer stands on. Skip what you already hold — most frontend engineers can move fast here, and several phases will be a review rather than a course.

Python as your second languageRead and write the language every AI tool is written in.0/5

FoundationPhase 1

Skip this if: You already ship Python services and are comfortable with typing, async, and packaging.

Skills to check off

Build this to finish the phase

Rewrite one small Node service in Python with FastAPI — same endpoints, same behavior.

Resources

No good course for this

  • Modern Python tooling — uv, ruff, mypy — has no course. Read the docs; coming from npm and tsc it is an afternoon.
SQL and databasesOwn the data layer, because retrieval quality decides AI quality.0/5

FoundationPhase 2

Skip this if: You write window functions and read query plans without looking things up.

Skills to check off

Build this to finish the phase

A schema with an embeddings column and working semantic search over your own data in Postgres.

Resources

No good course for this

  • Advanced SQL — query planning, window functions in depth — has no dedicated course in this list. Practice beats video here.
Data engineering and preparationGet data in, clean, and moving — half of AI work is this, unglamorously.0/5

FoundationPhase 3

Skills to check off

Build this to finish the phase

An ingest → clean → load pipeline that keeps a retrieval index up to date on a schedule.

Resources

No good course for this

  • Warehouses and the dbt/Snowflake/BigQuery world are a career of their own. Learn them only if your data actually lives there.
Backend and APIs for AIServe models behind an API that streams, retries, and does not fall over.0/4

FoundationPhase 4

Skip this if: You already design and operate production APIs — then take only the streaming and queueing parts.

Skills to check off

Build this to finish the phase

A service that streams model output to a React client and pushes long work to a queue.

Resources

Distributed systems and architectureReason about systems that fail partially, which is how AI systems always fail.0/4

FoundationPhase 5

Skills to check off

Build this to finish the phase

Split one LLM application into services with a queue and a cache in front.

Resources

No good course for this

  • Message queues have no hands-on course in this list — they appear inside system design courses. Build one with SQS or Redis and you will have learned more.
Cloud, containers, and deploymentShip and operate the thing, reliably and without surprise bills.0/4

FoundationPhase 6

Skip this if: You already own deployment and on-call for a production service.

Skills to check off

Build this to finish the phase

Containerize your project and deploy it with CI/CD and rollback.

Resources

Math and CS intuition for MLEnough intuition to reason about models — not a degree.0/4

FoundationPhase 7

Skills to check off

Build this to finish the phase

Compute cosine similarity between two embeddings by hand, then check it against a library.

Resources

Machine learning foundationsUnderstand the workflow LLM work inherited from classical ML.0/4

FoundationPhase 8

Skills to check off

Build this to finish the phase

Train and evaluate one scikit-learn model, from raw data to honest metrics.

Resources

No good course for this

  • PyTorch, which most current model work assumes, has no course here. Pick it up in phase 09 if you go that deep.

AI engineering core

0/46 skills

The model-specific depth that is the actual job. Nothing here is optional except the phase marked optional, and phase 07 is where senior candidates separate themselves.

The bridge into AI toolingGet productive in the AI toolchain using code you already understand.0/3

CorePhase 9

Skills to check off

Build this to finish the phase

A notebook that calls a model and returns a typed, schema-validated object.

Resources

How models actually workReason about why it failed instead of rearranging words until something sticks.0/5

CorePhase 10

Skills to check off

Build this to finish the phase

One prompt across three models and three temperatures — compare output, tokens, and cost.

Resources

Prompting, structured output, context engineeringMake model output reproducible enough to build a product on.0/5

CorePhase 11

Skills to check off

Build this to finish the phase

A versioned prompt library with typed outputs for three tasks, and test cases for each.

Resources

Your first real LLM applicationShip something people use before you touch any framework.0/4

CorePhase 12

Skills to check off

Build this to finish the phase

A full application with streaming, tool calls, and a designed failure state — deployed, not local.

Resources

RAG and vector searchGive a model private, current knowledge — reliably.0/5

CorePhase 13

Skills to check off

Build this to finish the phase

A RAG system on real data with hybrid retrieval, reranking, and measured retrieval quality.

Resources

No good course for this

  • No dedicated course exists on vector index types, hybrid search, or reranking in depth. Read your vector store's own documentation — that is where the real detail lives.
Agents, tools, and orchestrationMove from one call to systems that plan, act, and recover.0/5

CorePhase 14

Skills to check off

Build this to finish the phase

A stateful agent with three or more tools, memory, guardrails, and a replayable trace.

Resources

Evaluation and observabilityThe phase that separates senior candidates. Most people skip it.0/6

DifferentiatorPhase 15

Skills to check off

Build this to finish the phase

An eval harness with a golden set, scorers, and a CI gate — then a measured improvement to your project.

Resources

No good course for this

  • No standalone course exists on LLM observability — tracing, cost attribution, incident response. Vendor documentation and production experience are the teachers here.
Production, LLMOps, and securityKeep AI systems reliable, affordable, and safe once real users arrive.0/5

CorePhase 16

Skills to check off

Build this to finish the phase

Take your project to production with monitoring and a written threat model for prompt injection.

Resources

Open models and fine-tuningJudge when to leave the hosted API — and usually decide not to.0/4

OptionalPhase 17

Skills to check off

Build this to finish the phase

Fine-tune a small open model with LoRA and compare it against a frontier model on your own eval set.

Resources

No good course for this

  • Serving infrastructure — GPU scheduling, batching, model serving — is not covered anywhere in this path. That is where you hire or partner rather than self-teach.
AI system design and staying currentOperate at the level where you design the system and defend the trade-offs.0/4

DifferentiatorPhase 18

Skills to check off

Build this to finish the phase

Design and document a complete AI system — diagram plus a decision record — then get it reviewed by someone who will argue with you.

Resources

Concepts

What you should be able to explain out loud.

The interview surface, and the working vocabulary of the job. If you cannot explain one of these to a product manager, you have found your next study session.

Model mechanics

  • Tokens, tokenization, and why costs and limits are counted in them
  • Embeddings and vector similarity
  • Context windows, truncation, and context rot in long conversations
  • Temperature, top-p, and determinism (and its limits)
  • Hallucination: why it happens and which mitigations actually help
  • Base models vs instruction-tuned vs reasoning models

Prompting patterns

  • System, developer, and user roles, and what belongs in each
  • Few-shot examples and when they beat instructions
  • Chain-of-thought and structured reasoning
  • Structured output and schema validation
  • Prompt versioning, review, and regression testing
  • Prompt injection and untrusted content boundaries

Retrieval and context

  • Chunking strategies and why chunk boundaries decide answer quality
  • Embedding models and their trade-offs
  • Vector stores, index types, and hybrid search
  • Reranking and why the first retrieval pass is rarely enough
  • Metadata filtering — tenant, permission, recency
  • Cache strategy: prompt caching, embedding reuse, response caching

Agents and tools

  • Tool and function calling with schema validation
  • The agent loop: plan, act, observe, repeat — and its termination conditions
  • Memory: short-term, long-term, and what should never persist
  • Multi-agent orchestration and when it stops being worth it
  • MCP as an integration boundary
  • Human-in-the-loop checkpoints and approval flows

Evaluation and quality

  • Golden datasets and how to build one from real traffic
  • Deterministic scorers vs LLM-as-judge, and the bias in each
  • Offline evals vs online monitoring
  • Regression gates in CI for prompt and model changes
  • A/B testing generative features responsibly
  • Defining a quality bar with product owners before shipping

Production concerns

  • Cost per request, token budgets, and model routing by task
  • Latency budgets, streaming, and perceived performance
  • Rate limits, retries, backoff, and provider failover
  • Data privacy, PII handling, and retention in model calls
  • Authorization at the retrieval layer
  • Auditability: reconstructing why the system produced a given answer

System design

Design problems to practice.

Work these on paper until the trade-offs come naturally. They are the shape of both the interview and the actual job.

  1. 01Design a document assistant for an enterprise where every answer must cite approved internal sources and respect per-user permissions.
  2. 02Design a support agent that can act — issue refunds, change bookings — with a blast radius small enough for the business to approve.
  3. 03Design the evaluation and release pipeline for a generative feature, so a prompt change cannot silently regress quality in production.
  4. 04Design a retrieval system over data that changes hourly, and explain how the index stays fresh without reprocessing everything.
  5. 05Cut the cost of an existing AI feature substantially without a visible drop in quality — and prove the quality held.
  6. 06Design for provider outage and model deprecation: how does the product degrade rather than break?

Pace

How long this actually takes.

Honest expectations: the foundations track is a review for many experienced engineers and a real detour for others, and the AI core is roughly six months of consistent evening work — or three if you can make it part of your job, which is by far the fastest route. The resource hours add up to far more than that, but hours watched is the wrong metric. Ship the build for each phase, however small. The engineers who make this transition successfully have a deployed thing to point at, not a completed playlist.

  • Take the foundations you are missing, skip the ones you hold. Most frontend engineers need F1, F2, and F3 and can skim the rest.
  • Do core phases 01 through 04 in order — they compound, and skipping the model-mechanics phase is what produces prompt-shufflers.
  • Phase 05 through 07 is the actual job. Budget the most time there; phase 07 is what senior interviews probe hardest.
  • Phases 08 through 10 can be reordered around what your work or portfolio needs next.

FAQ

Questions people actually ask.

Do I need a machine learning degree or a maths background?

No. AI engineering is a software engineering discipline: you are integrating models into products, not training them. You need intuition, which the foundations track gives you in a few weeks, and then far more depth on data, retrieval, evaluation, and systems.

Can I stay in TypeScript and skip Python?

You can ship real products in TypeScript alone, and much of the current tooling is TypeScript-first. But you will be shut out of most research code, evaluation tooling, and data work, and you will struggle to collaborate with ML teams. Learn enough Python to read and modify it — that is weeks, not a career change.

Do I have to pay for any of this?

No. Every phase has at least one free resource, and the free options — fast.ai, Karpathy, Hugging Face, the Zoomcamps, the practitioner blogs — are genuinely competitive with the paid ones. Paid courses buy structure and sequencing, not exclusive knowledge.

How do I know I have finished a phase?

You built the thing. Every phase has one build, and that is the completion test — not the checkboxes and not the videos. If you cannot build it, the phase is not done, however much of it you have watched.

What should I build to prove I can do this?

One feature that a real person uses, with retrieval over data that matters, an eval harness showing you measured quality, and an honest writeup of what failed. That is worth more in an interview than any number of finished courses, because it is the only evidence that you can operate a non-deterministic system in production.

Does frontend experience actually count for anything here?

More than most people expect. Model APIs are slow, unreliable, streaming interfaces, and designing for partial and failed states is a frontend instinct. Most AI features fail as products before they fail as models — and product judgment is not something an ML background supplies.

Why I wrote this

I make this transition my daily work: I build AI product features at Writer — agent workflows, knowledge connectors, enterprise identity — on top of a frontend engineering background. This is the map I wish had existed, written from the inside rather than assembled from headlines.

Read: from AI prototype to productionMy engineering profile

Building an AI feature and stuck between demo and production?

That gap is where I work. Describe what is blocked and I will tell you whether I can help — and what the smallest useful next step would be.

Discuss a projector email js.leonid.kuznetsov@gmail.com